Fewshot Attempt 03

Credit Office Committee JSON Workflow

Generate committee-ready JSON answers from the shared credit office public API. Task prompts map to one of several template shapes; the skill below covers the universal data-gathering workflow, transferable computation rules, and output conventions.

1. Environment & Setup

• Base URL is given by the task prompt or GDPEVO_ENV_BASE_URL (e.g. http://34.46.77.124:8011).
• Never read local env/ source files; use only the public API.
• Start with:
  • GET /api/manifest – endpoint index, benchmark versions, record counts.
  • GET /api/policies – master policy (rating thresholds, CDFI factor tables, CRE weights, stress formulas, concentration rules).

2. Identify Task Type from Prompt

Read the prompt and input/payloads/answer_template.json before fetching data. Common shapes:

3. Data Fetching Order

1. Branch / segment details (/branches/{branch_id} or /credit-union-segments/{segment_id})
2. Metrics (/branches/{branch_id}/metrics) – use the quarter matching the review date (usually most recent).
3. Loans (/branches/{branch_id}/loans) – full portfolio for rating migrations / watch-list tasks.
4. Applications (/branches/{branch_id}/applications) – for allocation or competing-CRE tasks.
5. Sector exposures (/branches/{branch_id}/sector-exposures) – for concentration checks.
6. Benchmarks – banks: /benchmarks/fdic/q4-2024; credit unions: /benchmarks/ncua/q1-2025.

4. Transferable Computation Rules

Stress DSCR (CRE dual-stress)

• Formula: stressed_dscr = dscr * 0.85 / 1.18
• Breach threshold: 1.0
• Round base_dscr and stressed_dscr to 2 decimals.
• breaches_threshold is true when stressed_dscr < 1.0.

Watch-list stress (+200 bp)

• Formula: stressed_dscr = dscr / 1.18
• Same threshold and rounding.

CRE Concentration

• existing_cre_exposure = sum of outstanding_balance for all loans where loan_type == "CRE".
• existing_cre_concentration = existing_cre_exposure / total_loans_outstanding (from metrics), round to 4 decimals.
• selected_post_approval_cre_concentration = (existing_cre_exposure + selected_requested_amount) / (total_loans_outstanding + selected_requested_amount), round to 4 decimals.
• selected_policy_variance_bps = (selected_post_approval_cre_concentration - cre_policy_limit_pct) * 10000, round to 2 decimals.

FDIC Benchmark Variance (banks)

• branch_delinquency_ratio = delinquency_30_plus_pct from branch metrics.
• fdic_benchmark_ratio = total_real_estate_30_89_pct from FDIC Q4 2024.
• fdic_variance_ratio = branch_delinquency_ratio - fdic_benchmark_ratio, round to 4 decimals.
• fdic_variance_bps = fdic_variance_ratio * 10000, round to 2 decimals (e.g. 2802.0).

NCUA Variance (credit unions)

• branch_net_worth_ratio = segment-level net-worth ratio (from segment endpoint or derived from assets/deposits).
• ncua_benchmark_ratio = peer net_worth_ratio or state-level equivalent.
• ncua_variance_ratio = branch_net_worth_ratio - ncua_benchmark_ratio.
• ncua_variance_bps = ncua_variance_ratio * 10000.

Risk Rating Re-derivation (dominant-factor rule)

1. DSCR rating: lookup from policy dscr_thresholds.
2. LTV rating: lookup from policy ltv_thresholds.
3. Delinquency rating: lookup from policy delinquency_minimums.
4. Final rating = worst (highest number) of the available factors.
5. Downgrade notches = final_rating - current_rating.

CDFI Factor Score → Risk Class

• Sum factor scores from policy tables (fico, ltv, debt_to_asset, liquidity_months).
• Map total to class:
  • 0–5 → Prime
  • 6–9 → Desirable
  • 10–13 → Satisfactory
  • 14–18 → Watch
  • >=19 → Doubtful
  • >=19 and ltv > 1.0 → Projected Loss

CRE Weighted Score → Score Class

• weighted_cdfi_score is computed from the 5 CDFI factors weighted by policy (capacity 0.45, capital 0.03, character 0.05, collateral_exposure 0.36, conditions 0.11).
• Map to class:
  • <= 2.0 → approve_quality
  • <= 3.0 → conditional
  • > 3.0 → weak

5. Output Field Conventions

• JSON only — no markdown, no narrative outside the JSON object.
• Precision:
  • Currency / exposure: 2 decimals.
  • Percentages as ratios: 4 decimals.
  • BPS values: 2 decimals.
  • DSCR: 2 decimals.
  • Weighted CDFI score: 1 decimal.
• Ordering:
  • Lists of IDs (loan_id, application_id): ascending alphanumeric.
  • Reason codes / conditions / enums: ascending alphabetical.
  • Concentration flags: by sector, then application_id.
  • Workout queue: descending exposure, then ascending loan_id.
  • Severe bucket counts: ascending current_rating, then payment_status.
• Enums: use exact allowed values from the template (case-sensitive). Never invent new values.

6. Decision & Reason-Code Heuristics

• Approve quality (<=2.0) → usually approve.
• Conditional (2.0–3.0) → conditional_approve or participation_required if concentration/sector limits are breached.
• Weak (>3.0) → defer or decline.
• Common reason codes:
  • sector_breach – post-approval sector % > sector_ceiling_pct.
  • weak_dscr – stressed DSCR breaches threshold.
  • high_ltv – LTV exceeds policy thresholds.
  • fdic_adverse_variance – branch delinquency materially exceeds FDIC benchmark.
  • capacity_limit – approval would exceed lending_capacity_q1.
  • documentation_gap – documentation_complete == 0.
• Unselected application in a competing-CRE task receives the disposition of the weaker credit (decline or defer) and its reason codes (alphabetically sorted).

7. Common Pitfalls

• Forgetting the template: Always read input/payloads/answer_template.json first; required keys and enum choices differ by task.
• Wrong benchmark set: Banks use FDIC Q4 2024 (total_real_estate_30_89_pct); credit unions use NCUA Q1 2025 state rows.
• Wrong denominator: CRE concentration uses total_loans_outstanding from branch metrics, not total_assets.
• Off-by-one in ratings: Delinquency minimums are floors — if a loan is 90+ Days Past Due, its rating cannot be better than 7 regardless of DSCR/LTV.
• Missing loans in watch-list / migration: Include all loans that meet the threshold (e.g., current_rating >= 3 for migration, current_rating >= 6 for watch-list), not just a subset.
• NPA calculation: Non-performing loans are those with payment_status of 90+ Days Past Due or Nonaccrual, or current_rating in the severe bucket. branch_npa_ratio = npa_exposure / total_loans_outstanding.
• Alphabetical sorting: Python's default string sort works; verify that reason codes like ["fdic_adverse_variance", "sector_breach"] are correctly ordered.